Use this URL to cite or link to this record in EThOS: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.640969
Title: 3D computer vision : passive depth from defocus and manifold learning-based human activity recognition
Author: Li, Ang
Awarding Body: University of Warwick
Current Institution: University of Warwick
Date of Award: 2014
Availability of Full Text:
Access from EThOS:
Access from Institution:
Abstract:
The rational operator-based approach to depth from defocus (DfD) using pill-box point spread function (PSF) enables texture-invariant 3-dimensional (3D) surface reconstructions. However, pill-box PSF produces errors when the amount of lens diffraction and aberrations varies. This thesis proposes two DfD methods, one using the Gaussian PSF that addresses the situation when diffraction and aberrations are dominant, and the second based on the generalised Gaussian PSF that deals with any levels of the problem. The accuracy of DfD can be severely reduced by elliptical lens distortion. This thesis also presents two correction methods, correction by distortion cancellation and correction by least squares fit. Each method is followed by a smoothing algorithm to address the low-texture problem of DfD. Most existing human activity recognition systems pay little attention to an effective way to obtain training silhouettes. This thesis presents an algorithm to obtain silhouettes from any view using 3D data produced by Vicon Nexus. Existing background subtraction algorithms produce moving shadow that has a significant impact on silhouette-based recognition system. Shadow removal methods based on colour and texture fail when the surrounding background has similar colour or texture. This thesis proposes an algorithm based on known position of the sun to remove shadow in outdoor environment, which is able to remove essential part of the shadow to suffice recognition purpose. Unlike most recognition systems that are either speed-variant, temporal-order-variant, inefficient or computational expensive, this thesis presents a near real-time system based on embedded silhouettes. Silhouettes are first embedded with isometric feature mapping, and the transformation is learned by radial basis function. Complex human activities are then learnt with spatial objects created from the patterns of embedded silhouettes.
Supervisor: Not available Sponsor: School of Engineering, University of Warwick
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.640969  DOI: Not available
Keywords: TA Engineering (General). Civil engineering (General)
Share: